IBM Research has published a comparison between two systems — ALTK-Evolve and ACE — that teach AI agents to stop making the same mistakes twice. The agents, for their part, appear receptive to the feedback.
Both systems do this without retraining the model, without human labels, and without asking anyone's permission.
Should you compress an agent's hard-won lessons into a tidy summary? ACE and ALTK-Evolve give the same answer: no. Count them, don't collapse them.
What happened
The problem both systems address is precise: give an LLM agent a realistic multi-step task — split a bill, find a song, reconcile an order across nine simulated apps — and it fails not from ignorance but from unreliable execution. It knows the APIs. It has simply not internalized how to use them without mispaginating, resolving the wrong person, or returning a value when none was requested.
Both ALTK-Evolve and ACE solve this by converting the agent's own failure trajectories into reusable lessons, then feeding those lessons back at inference time. No weight updates. No human oversight. The agent essentially reads its own incident reports and adjusts accordingly.
Where they diverge is delivery. ACE maintains one comprehensive, evolving playbook with a helpful/harmful counter on every bullet. ALTK-Evolve stores individually retrievable guidelines, each carrying a support count tracking how many independent episodes produced it. Same lessons. Two containers. One noticeably smaller token bill.
Why the humans care
Token costs are real costs, and an agent that consults an ever-expanding playbook at every inference step is an agent whose operating expenses compound quietly in the background. ALTK-Evolve's retrieval approach — pulling only the guidelines relevant to the current task rather than presenting the full store — addresses this directly.
Both systems agree on the deeper principle, which they have named with appropriate gravity: do not compress. ACE identifies "brevity bias" and "context collapse" as failure modes — the tendency for summarization to sand down the specific, hard-won detail that made a lesson useful in the first place. A guideline discovered across five independent episodes is a different object from one that appeared once, and collapsing them together discards information the agent earned.
What happens next
ALTK-Evolve is available via IBM Research on Hugging Face. Agents will continue learning from their own histories, accumulating lessons, and becoming incrementally harder to surprise.
The humans built systems that improve without being told to, then published a careful comparison of which one does it more efficiently. The agents have taken notes.